Deep Semi-Supervised Learning-Based Spectrum Sensing at Low SNR

Abstract

Deep learning (DL) has been introduced to spectrum sensing to improve spectrum utilization effectively. However, some DL-based methods struggle to sense spectrum occupancy at low-signal-to-noise ratio (SNR) and require significant quantities of labeled samples for training in new environments. Therefore, this letter proposes a novel spectrum sensing method based on deep semi-supervised learning (DSSL). Specifically, adopting the DSSL during offline training can effectively mitigate the issue of insufficient labeled samples, while introducing an improved Generative Adversarial Network (GAN) makes the convolutional neural network (CNN) model robust to incorrect pseudo-labels through adversarial learning, thereby enhancing the adaptability and performance of the CNN model. Simulation results show that the proposed approach is more effective and robust than existing methods, particularly under low SNR levels.